Under-10-Person Chinese Team's 27B Security Model Hits 63% on CyberGym, Near Claude
新智元 · wechat · 2026-09-16
The story
VeraPraxis, a Chinese startup with fewer than 10 people, released Feyospace-v1, a 27B cybersecurity model that scored 63.0% on CyberGym's Model-focused leaderboard — approaching Claude Sonnet 4.6's 65.2%. CyberGym tests real-world vulnerability reproduction across 1,507 tasks from 188 software projects, requiring models to analyze code, craft triggering tests, and iterate on execution feedback.
Technical highlights
- One post-training recipe improved three different base models; the 27B base gained 8.73 points
- On ExploitBench, it lifted a Qwen 8B-class model from 12.96% to 19.05%
- Kreator method: expert insights at key decision points are turned into full learnable reasoning-action traces, preserving intermediate derivations so smaller models learn the process, not just surface answers
- 164k training trajectories collected via executable environments with execution verification and evidence auditing
Company strategy
VeraPraxis runs four businesses (apps, audio — HeartMuLa music model open-sourced, security, quant) as a flywheel: real operations surface problems models can't solve, experts fill the gap, and its PraxisOS organizes goals, actions, expert feedback, and outcomes into a continuous model-development loop, with RL planned next.
More from Companies & People
- ElevenLabs in one week: UMusic deal, Music v2.5, MCP multimodal, $1M AI SDR — lukeharries · 2026-09-16
- NUS MAGIC Lab is hiring research assistants and engineers in robotics and post-training — DJiafei · 2026-09-16
- Retrieval researcher fuma_nama joins Mixedbread to build information retrieval for AI agents — bclavie · 2026-09-16
- Ex-Pluralsight exec on why climbing the ladder was 'a sophisticated form of hiding' — JnBrymn · 2026-09-16
- Base Power, valued at $13B, hosts 48-hour hiring hackathon in Austin — lifebypixels · 2026-09-16
- AI momentum is a melting ice ball: durability is a myth, says Stanford's Ann Bordetsky — annbordetsky · 2026-09-16